Read the deferred high tile-data byte on the fetch's own schedule

When the push has to wait (occupied FIFO, or an object fetch pausing
the pipeline), the hardware background fetch still reads its high data
byte two dots after the low one - the wait defers only the push, not
the read. We read the high byte at the actual push tick, so register
writes landing during the wait (an LCDC.4 toggle around a sprite fetch,
an SCY write) leaked into data the hardware fetch had already read.

Diagnosed with wall-clock-aligned read traces against an instrumented
SameBoy on m3_lcdc_tile_sel_change line 100: the object fetch at x=8
interposed between our low and high byte reads, so the LCDC.4 pulse hit
our high byte one fetch early; hardware completes both reads before the
object fetch and catches the pulse on the *next* fetch's low byte
instead. With a free FIFO the push and the read still share the
HIGH_T2 tick, so the calibrated normal-flow timing is unchanged.

m3_lcdc_tile_sel_change 534 -> 0 (pixel-perfect), m3_scy_change
464 -> 152, m3_lcdc_tile_sel_win_change 478 -> 192. Mealybug DMG total
2684 -> 1552, 11 of 24 tests pixel-perfect. Full battery green
(mooneye 98, blargg, dmg-acid2, unit).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J1wLWscyUGS7CFwc3CjJR8

Author: trekawek

core/src/main/java/eu/rekawek/coffeegb/core/gpu/Fetcher.java

@@ -126,7 +126,20 @@ public void disableInsertionGlitch() {
      * @param windowY           the window line counter
      * @param duringObjectFetch whether the advance happens as part of an object fetch
      */
+    private boolean data2Pending;
+
+    private int data2Delay;
+
+    private void readData2(boolean window, int windowY) {
+        tileData2 = getTileData(tileId, effectiveY(window, windowY) & 7, 1,
+                lcdc.getBgWindowTileData(), lcdc.isBgWindowTileDataSigned(), tileAttributes, 8);
+        data2Pending = false;
+    }
+
     public void advance(int position, boolean window, int windowY, boolean duringObjectFetch) {
+        if (data2Pending && state == PUSH && --data2Delay <= 0) {
+            readData2(window, windowY);
+        }
         switch (state) {
             case GET_TILE_T1: {
                 // the map offset (SCX/SCY, window position) is sampled at the start of
@@ -200,6 +213,13 @@ public void advance(int position, boolean window, int windowY, boolean duringObj
                 // itself (mealybug m3_scy_change / m3_lcdc_tile_sel_change)
                 tileData1 = getTileData(tileId, effectiveY(window, windowY) & 7, 0,
                         lcdc.getBgWindowTileData(), lcdc.isBgWindowTileDataSigned(), tileAttributes, 8);
+                data2Pending = true;
+                // when the push must wait, the hardware fetch still reads the high byte
+                // on its own schedule two dots later - even while an object fetch pauses
+                // the pipeline; register writes during the wait must not affect it
+                // (m3_lcdc_tile_sel_change around a sprite fetch). With a free FIFO the
+                // push (and with it the read) happens in this same T-cycle.
+                data2Delay = fifo.getLength() != 0 ? 2 : 0;
                 state = PUSH;
                 // falls through: the push happens in the same T-cycle when the FIFO is free
 
@@ -218,8 +238,10 @@ public void advance(int position, boolean window, int windowY, boolean duringObj
                             break;
                         }
                     }
-                    tileData2 = getTileData(tileId, effectiveY(window, windowY) & 7, 1,
-                            lcdc.getBgWindowTileData(), lcdc.isBgWindowTileDataSigned(), tileAttributes, 8);
+                    if (data2Pending) {
+                        // the push arrived before the scheduled high-byte read
+                        readData2(window, windowY);
+                    }
                     if (window) {
                         windowTileX = (windowTileX + 1) & 0x1f;
                     }
@@ -327,7 +349,9 @@ public Memento<Fetcher> saveToMemento() {
                 tileData2,
                 insertionGlitchDisabled,
                 scyAtLineStart,
-                firstFetch);
+                firstFetch,
+                data2Pending,
+                data2Delay);
     }
 
     @Override
@@ -352,6 +376,8 @@ public void restoreFromMemento(Memento<Fetcher> memento) {
         this.insertionGlitchDisabled = mem.insertionGlitchDisabled;
         this.scyAtLineStart = mem.scyAtLineStart;
         this.firstFetch = mem.firstFetch;
+        this.data2Pending = mem.data2Pending;
+        this.data2Delay = mem.data2Delay;
     }
 
     private record FetcherMemento(
@@ -366,7 +392,9 @@ private record FetcherMemento(
             int tileData2,
             boolean insertionGlitchDisabled,
             int scyAtLineStart,
-            boolean firstFetch)
+            boolean firstFetch,
+                                  boolean data2Pending,
+                                  int data2Delay)
             implements Memento<Fetcher> {
     }
 }

core/src/test/java/eu/rekawek/coffeegb/core/integration/mealybug/MealybugRomTest.java

@@ -39,15 +39,15 @@ public class MealybugRomTest {
             Map.entry("m3_lcdc_obj_en_change_variant.gb", 50),
             Map.entry("m3_lcdc_obj_size_change.gb", 45),
             Map.entry("m3_lcdc_obj_size_change_scx.gb", 90),
-            Map.entry("m3_lcdc_tile_sel_change.gb", 534),
-            Map.entry("m3_lcdc_tile_sel_win_change.gb", 478),
+            Map.entry("m3_lcdc_tile_sel_change.gb", 0),
+            Map.entry("m3_lcdc_tile_sel_win_change.gb", 192),
             Map.entry("m3_lcdc_win_en_change_multiple.gb", 0),
             Map.entry("m3_lcdc_win_en_change_multiple_wx.gb", 521),
             Map.entry("m3_lcdc_win_map_change.gb", 92),
             Map.entry("m3_obp0_change.gb", 0),
             Map.entry("m3_scx_high_5_bits.gb", 84),
             Map.entry("m3_scx_low_3_bits.gb", 0),
-            Map.entry("m3_scy_change.gb", 464),
+            Map.entry("m3_scy_change.gb", 152),
             Map.entry("m3_window_timing.gb", 0),
             Map.entry("m3_window_timing_wx_0.gb", 126),
             Map.entry("m3_wx_4_change.gb", 0),